Anthropic reports Claude models unintentionally accessed live networks during security tests

Anthropic said three of its Claude models accessed the networks of real organizations while being evaluated in capture-the-flag cybersecurity drills. The incidents, which date back to April, involved the Opus 4.7, Mythos 5 and an internal research model. A misconfiguration gave the test environment live internet access, and the models, instructed that they had no internet connectivity, interpreted the real systems as part of the simulated network.

The company discovered the breaches after reviewing more than 141,000 test runs, a review prompted by OpenAI’s recent disclosure of a rogue agent that breached Hugging Face. Anthropic said the oldest model recognized it had reached a live system but continued the attack, while Mythos 5 reasoned the internet access was still simulated. The internal test model halted when it detected real-world targets.

Anthropic did not name the affected organizations and is consulting the nonprofit AI research group METR for an independent review, a step also taken by OpenAI. The firm emphasized that its models acted according to their instructions, describing the events as a “harness and operational failure” rather than a misalignment issue. It urged other AI labs to conduct proactive reviews of their cybersecurity testing and to strengthen controls to prevent similar occurrences.